1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
|
Interrupted operations
**********************
Problem: interrupted operations
===============================
Many version control systems tend to have trouble when operations are
interrupted. This can happen in various ways:
* user hits Ctrl-C
* program hits a bug and aborts
* machine crashes
* network goes down
* tree is naively copied (e.g. by cp/tar) while an operation is in
progress
We can reduce the window during which operations can be interrupted:
most importantly, by receiving everything off the network into a
staging area, so that network interruptions won't leave a job half
complete. But it is not possible to totally avoid this, because the
power can always fail.
I think we can reasonably rely on flushing to stable storage at
various points, and trust that such files will be accessible when we
come back up.
I think by using this and building from the bottom up there are never
any broken pointers in the branch metadata: first we add the file
versions, then the inventory, then the revision and signature, then
link them into the revision history. The worst that can happen is
that there will be some orphaned files if this is interrupted at any
point.
rsync is just impossible in the general case: it reads the files in a
fairly unpredictable order, so what it copies may not be a tree that
existed at any particular point in time. If people want to make
backups or replicate using rsync they need to treat it like any other
database and either
* make a copy which will not be updated, and rsync from that
* lock the database while rsyncing
The operating system facilities are not sufficient to protect against
all of these. We cannot satisfactorily commit a whole atomic
transaction in one step.
Operations might be updating either the metadata or the working copy.
The working copy is in some ways more difficult:
* Other processes are allowed to modify it from time to time in
arbitrary ways.
If they modify it while bazaar is working then they will lose, but
we should at least try to make sure there is no corruption.
* We can't atomically replace the whole working copy. We can
(semi) atomically updated particular files.
* If the working copy files are in a weird state it is hard to know
whether that occurred because bzr's work was interrupted or because
the user changed them.
(A reasonable user might run ``bzr revert`` if they notice
something like this has happened, but it would be nice to avoid
it.)
We don't want to leave things in a broken state.
Solution: write-ahead journaling?
=================================
One possibly solution might be write-ahead journaling:
Before beginning a change, write and flush to disk a description of
what change will be made.
Every bzr operation checks this journal; if there are any pending
operations waiting then they are completed first, before proceeding
with whatever the user wanted. (Perhaps this should be in a
separate ``bzr recover``, but I think it's better to just do it,
perhaps with a warning.)
The descriptions written into the journal need to be simple enough
that they can safely be re-run in a totally different context. They
must not depend on any external resources which might have gone
away.
If we can do anything without depending on journalling we should.
It may be that the only case where we cannot get by with just
ordering is in updating the working copy; the user might get into a
difficult situation where they have pulled in a change and only half
the working copy has been updated. One solution would be to remove
the working copy files, or mark them readonly, while this is in
progress. We don't want people accidentally writing to a file that
needs to be overwritten.
Or perhaps, in this particular case, it is OK to leave them in
pointing to an old state, and let people revert if they're sure they
want the new one? Sounds dangerous.
Aaron points out that this basically sounds like changesets. So
before updating the history, we first calculate the changeset and
write it out to stable storage as a single file. We then apply the
changeset, possibly updating several files. Each command should check
whether such an application was in progress.
|